Mantis: Fast and Communication-Friendly Secure Two-Party Neural Network Inference

Secure two-party (2PC) neural network inference allows a client to classify images by using the server’s model while protecting the privacy of both parties. Although many of the state-of-the-art schemes are fast and communication-friendly, the overhead of their online phase remains considerable. Function secret sharing (FSS) based techniques offer an attractive solution to this in the trusted dealer model, and 2PC protocols based on these techniques have a very lightweight online phase for non-linear layers. However, existing FSS-based solutions still have large communication, computation, and storage overhead for linear layers. In this work, we propose Mantis, a new semi-honest secure 2PC-NN inference system that is faster and more communication-friendly in the online phase than state-of-the-arts. In the proposed system, due to the feature that the convolutional kernel matrix is plaintext for the server, we present an efficient and secure 2PC matrix convolution protocol Fast MatrixConv by leveraging some random tuples related to the kernel matrix generated in the offline phase, which only a small amount of computation, communication and storage are required in the online phase. Furthermore, according to the fact that the size of the convolutional kernel matrix is much smaller than the size of the image matrix, we perform batch processing operations in the offline stage by using VOLE technique and OT stitching technique, which greatly reduces the overhead of the preprocessing operation in the offline stage in Fast MatrixConv. Compared to the state-of-the-art work LLAMA, Mantis can reduce communication by 11.45% and runtime by 10.75% under ResNet50 inference in the online phase.